How long does an AI citation last? Six weekly runs across six engines

How we measured it

The questions are the 40 fixed English buyer questions of our AI Source Index: 16 "best X" questions, 8 "alternatives to X", 8 head-to-head comparisons and 8 "how do I choose" questions. Every week they go to six answer engines through their APIs: Perplexity, ChatGPT (OpenAI's API with web search, gpt-4o-mini), Gemini, Claude, Google AI Overviews and Google AI Mode. We used every run the index has published so far: 3, 7, 14, 21 and 28 September and 5 October 2026. Each engine is asked each question once per run.

• Unit: one domain in one answer. A domain is the lowercase host with a leading www. removed, counted once per answer however many of its pages were cited. Google's own shopping links are dropped, as the public index drops them.
• Retained: cited again for the same question by the same engine in the next run.
• Failed checks are left out, not counted as losses. 29 of Claude's 40 checks failed on 14 September and one AI Overviews check failed on 5 October; those question-engine pairs leave every comparison that touches those runs. An answer that came back with no source at all stays in: if a domain was cited one week and the next answer cited nothing, it counts as gone.
• Every rate is a count out of a count with a 95% Wilson interval. The same domains recur within a question, so those intervals are slightly too narrow; resampling whole questions gives 46.2–49.8% for the headline rate instead of 46.8–49.2%.

The rows behind every number are public in the AI Source Index, one line per question, engine and run, and we are publishing the analysis script beside them. How the index collects its answers is on the methodology page.

About half of cited domains are back the next week

Pooled over the five week-to-week pairs, 3,335 of the 6,947 domains an engine cited for a question were cited again for the same question by the same engine one run later: 48.0%. The individual pairs ranged from 42.4% (28 September to 5 October, 625 of 1,474) to 55.2% (3 to 7 September, 782 of 1,416). That highest pair is also the only one four days apart instead of seven.

Pages turn over slightly faster than domains. The same URL was cited again the next week 41.4% of the time (3,201 of 7,739). But when a domain stayed, at least one of the same pages came with it in 90.9% of cases (3,032 of 3,335). Week to week, what changes is mostly which sites an answer draws on, not which page of a site it keeps.

Most new citations are gone a week later, and old ones flicker

Follow the citations that appeared for the first time in one of the runs, and most are gone a week later:

• Back the next week: 533 of 1,950, 27.3%.
• Present in each of the next two weeks: 193 of 1,497, 12.9%.
• Each of the next three: 69 of 1,058, 6.5%.
• Each of the next four: 26 of 568, 4.6%.

Survival in an unbroken run understates how many sources come back, though. Of the domains cited on 3 September, only 11.0% were cited in every one of the next five runs (139 of 1,259), yet 36.4% were cited again on 5 October (458 of 1,259). Many of them were missing for a week or two in between.

Over all six runs, half of the distinct combinations of question, engine and domain appeared in only one run: 1,781 of 3,592, 49.6%. Only 139 (3.9%) appeared in all six. The stable ones are counted again every week they appear, so the one-run citations make up a smaller share of everything an engine cited: 1,781 of 7,644 citations, 23.3%.

Staying power builds. A first-time citation made the next week 27.3% of the time; one that had already held for two runs in a row made a third 44.5% of the time (193 of 434), and after three in a row, 47.3% (69 of 146).

Decay or noise? Mostly reshuffling, with some drift

Each engine answers each question once a week, and these engines do not give the same answer every time they are asked. So a domain missing this week may be the engine's own randomness rather than a decision that the page no longer deserves a citation. We do not ask twice within a week, so we cannot measure that noise directly. Two things in the data point to it.

• Dropped domains come back. Of the domains that dropped out between two runs, 25.3% were back in the very next run (682 of 2,697) and 39.6% were cited again in some later run of the period (1,092 of 2,761).
• The loss does not keep growing at the same pace. 48.0% of domains were cited again one run later, 42.5% two runs later, 42.7% three, 37.5% four and 37.4% five (530 of 1,416, 3 September to 5 October). If each week's answers were random draws from a fixed set of sources, that curve would be flat; it falls about 10 points after the first step, so some sources really do fade. But most of the turnover happens in the first week and does not compound.

The honest reading: a large part of week-to-week churn is the engines reshuffling a pool of sources they keep returning to, and a smaller part is genuine decay. Six runs cannot say how long the pool itself lasts beyond five weeks.

Engines differ, but one week cannot rank them

Pooled over all five pairs, Claude and Perplexity kept the most of their sources and Google AI Mode the fewest:

• Claude: 563 of 1,033, 54.5% (fewer questions, because of the 14 September failures).
• Perplexity: 537 of 1,000, 53.7%.
• ChatGPT: 611 of 1,300, 47.0%.
• Gemini: 937 of 2,029, 46.2%.
• Google AI Overviews: 502 of 1,123, 44.7%.
• Google AI Mode: 185 of 462, 40.0%.

Resampling whole questions, the intervals for Claude (48.0–60.7%) and Perplexity (50.6–56.4%) do not overlap AI Mode's (32.9–46.9%). ChatGPT, Gemini and AI Overviews cannot be told apart from each other. And no single week pair gives this order. Perplexity kept 65.3% of its sources from 3 to 7 September and 33.2% from 21 to 28 September; AI Overviews went from 59.9% to 26.6% over the same two pairs.

We made that mistake ourselves. On 5 October we quoted one week pair in public, 28 September to 5 October, in which Gemini looked the most stable engine at 53% and the overall figure was 44%. Pooled over five pairs, Gemini is fourth. That count was taken from our database, where Google's own shopping links still count as a domain, and it skipped questions where either answer cited nothing. On the public rows and the rules in this guide, the same pair is 41.4%.

Part of AI Mode's low number is not switching sources at all. In each run, 10 to 16 of its 40 answers cited no source outside Google itself. Counting only pairs where both answers cited something, AI Mode kept 49.1% of its domains (185 of 377).

The biggest sources are the stickiest

youtube.com and reddit.com were the two most cited domains in all six runs, and 7 to 9 of each week's ten most cited domains were still in the top ten the next week. The domains that were already big held on better: the 20 most cited domains of 3 September were retained 60.0% of the time in the later week pairs (604 of 1,007), against 43.1% for every other domain (1,949 of 4,524).

For a smaller site, a single citation is weak evidence of anything. A citation that has held for three or four consecutive weeks is a different thing from one that showed up once.

How this compares with Profound's and Writesonic's citation decay

These answer a different question from ours. They follow a page's share of all citations as it rises and falls, day by day, across whatever questions they track. We ask whether the same engine cites the same domain for the same fixed question a week later. Both find most citations changing within a few weeks: Profound's median page halves its share 11 days after its peak, and in our panel about half of the cited domains were not cited again the following week. Their samples are far larger than our 40 questions, and our weekly runs cannot see a page that rises and fades inside one week. What this study adds is a fixed question set, an interval on every rate, and rows anyone can recount.

What to do with this

• Do not react to one week. A domain that dropped out came back the following week a quarter of the time in our data.
• Judge a citation by how long it has held. In our data, one that had held for two runs in a row made a third 44.5% of the time; a first appearance made a second only 27.3% of the time.
• Compare engines over several weeks, never on a single pair of runs. The order changed from one week pair to the next.
• Track the domain first. When the domain stayed in an answer, the same page came with it nine times in ten.
• When a Google AI Mode citation disappears, check whether the new answer cited anything at all before concluding a competing page replaced yours.

For which kinds of sites the engines cite in the first place, see what AI engines actually cite. For how Google's AI answers relate to its own first page, see Google's AI cites pages Google doesn't rank.

Frequently asked questions